SORT: SIMPLE ONLINE AND REALTIME TRACKING
Date: 2017
Journal: CVPR
1. Introduction
The MOT problem can be viewed as a data association problem where the aim is to associate detection across the frames
There is a resurgence of mature data association techniques including Multiple Hypothesis Tracking(MHT) and Joint Probabilistic Data Association(JPDA) which occupy many of the top positions of the MOT benchmark
Traditional Tracker is too slow for realtime applications
Instead of focusing on efficient and reliable handling of the common frame to frame associations, exploit recent advances in visual object detection to solve detection problem directly
2. Literature review
Traditional MOT delay making difficult decisions while there is high uncertainty over the object assignments
Many online tracking methods aim to build appearance models of either the individual objects themselves or a global model through online learning
When considering only one-to-one correspondence modelled as bipartite graph matching, globally optimal solutions such as the Hungarian algorithm can be used
3. Methodology
3.1 Detection
Utilize the Faster Region CNN (FrCNN) detection framework, which is an end to end framework that consist of two stages in this paper
first stage extracts features and proposes region, second stage classifies
Can be swapped to any design
3.2 Estimation Model
The inter-frame displacements of each object with a linear constant velocity model which is independent of other objects and camera motion
When a detection is associated to target. the bounding box is used to update the target state where the velocity components are solved optimally via Kalman filter framework
If no detection is associated to the target, its state is simply predicted without correction using the linear velocity model
3.3 Data Association
The assignment cost matrix is the computed as the intersection over union distance between each detection and all predicted bounding boxes from the existing targets
The assignment is solved optimally using Hungarian algorithm
3.4 Creation and deletion of Track Identities
For any detection with an overlap less than to signify the existence of an untracked object
Tracks are terminated if they are not detected for frames to prevent an unbounded growth in the # of trackers and localisation errors
Small cause early deletion of lost targets which aids eddiciency
